A deep learning system for non-invasive breast cancer diagnosis with multimodal data - PubMed
5 hours ago
- #breast cancer diagnosis
- #deep learning
- #multimodal imaging
- BINDS is a deep learning system integrating multimodal medical imaging for non-invasive breast cancer risk assessment and subtype classification.
- The system employs a two-stage approach aligning with clinical workflows: initial screening with ultrasound/mammography followed by multimodal diagnosis incorporating MRI.
- A novel radiology-pathology alignment mechanism is introduced to extract pathology-relevant features from radiological images.
- BINDS was validated on a diverse dataset of 27,048 participants from 8 centers and 7 public datasets, supporting flexible input modality combinations.
- Performance: Achieved an AUC of 0.973 and can reduce biopsies for benign lesions by up to 32.4%, aiding radiologists in precise decision-making.